Response Surface in low-rank Tensor Train Format for Uncertainty Quantification Alexander Litvinenko 1 (joint work with Sergey Dolgov 2 , Boris Khoromsij 2 and Hermann G. Matthies 3 ) 1 KAUST, 2 Max-Planck-Institut f ¨ ur Mathematik in den Naturwissenschaften, Leipzig, 3 Technische Universit¨ at Braunschweig, Germany Center for Uncertainty Quantification http://sri-uq.kaust.edu.sa/
41
Embed
Response Surface in low-rank Tensor Train Format for Uncertainty … · 2020. 3. 14. · Response Surface in low-rank Tensor Train Format for Uncertainty Quantification Alexander
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Response Surface in low-rank Tensor TrainFormat for Uncertainty Quantification
Alexander Litvinenko1
(joint work with Sergey Dolgov2, Boris Khoromsij2 andHermann G. Matthies3)
1 KAUST, 2 Max-Planck-Institut fur Mathematik in denNaturwissenschaften, Leipzig,
3 Technische Universitat Braunschweig, Germany
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
1. Parameter q := q(ω) (assume moments/cdf/pdf/quantilesof q are given)
2. Boundary and initial conditions, right-hand side3. Geometry of the domain
Uncertain solution:1. mean value and variance of u2. exceedance probabilities P(u > u∗)3. probability density functions (pdf) of u.
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
2 / 41
4*
Motivation
Nowadays computational algorithms, run onsupercomputers, can simulate and resolve verycomplex phenomena. But how reliable are thesepredictions? Can we trust to these results?
Some parameters/coefficients are unknown,lack of data, very few measurements→uncertainty.
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
3 / 41
4*
Example: Realisations of random fields
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
4 / 41
What is Quantification of uncertainties ?A big example:
UQ in numerical aerodynamics(described by Navier-Stokes + turbulence modeling)
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
5 / 41
4*
Example: uncertainties in free stream turbulence
α
v
v
u
u’
α’
v1
2
Random vectors v1(θ) and v2(θ) model free stream turbulence
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
6 / 41
4*
Example: UQ
Input parameters: assume that RVs α and Ma are Gaussianwith
mean st. dev.σ
σ/mean
α 2.79 0.1 0.036Ma 0.734 0.005 0.007
Then uncertainties in the solution (lift force and drag force) are
lift force 0.853 0.0174 0.02drag force 0.0206 0.003 0.146
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
7 / 41
4*
500 MC realisations of the pressure in dependence on αi and Mai
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
8 / 41
4*
Consider diffusion equation with uncertain coeffs
− div(κ(x , ω)∇u(x , ω)) = p(x , ω) in G × Ω, G ⊂ R3,u = 0 on ∂G, (1)
where κ(x , ω) - conductivity coefficient. Since κ positive,usually κ(x , ω) = eγ(x ,ω).
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
9 / 41
Discretisation of stochastic PDE
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
10 / 41
4*
Karhunen-Loeve Expansion
The Karhunen-Loeve expansion is the series
κ(x , ω) = µk (x) +∞∑
i=1
√λiki(x)ξi(ω), where
ξi(ω) are uncorrelated random variables and ki are basisfunctions in L2(G).Eigenpairs λi , ki are the solution of
Tki = λiki , ki ∈ L2(G), i ∈ N, where.
T : L2(G)→ L2(G),(Tu)(x) :=
∫G covk (x , y)u(y)dy .
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
11 / 41
4*
KLE eigenfunctions in 2D
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
12 / 41
4*
Problem with Polynomial Chaos Expansion
ξi(ω) ≈Z∑
k=0
ak Ψk (θ1, θ2, ..., θM),
where Z = (M+p)!M!p! or Z = pM :
- EXPENSIVE!M = 9, p = 2, Z = 55M = 9, p = 4, Z = 715M = 100, p = 4, Z ≈ 4 · 106.How to store and to handle so many coefficients ?The orthogonality of Ψk enables the evaluation
ak =< ξΨk >
< Ψ2k >
=1
< Ψ2k >
∫ξ(θ(ω))Ψk (θ(ω))dP(ω).
(e.g. Ψk are multivariate Hermite polynomials).Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
13 / 41
4*
Smooth transformation of Gaussian RF
We assume κ = φ(γ) -a smooth transformation of the Gaussianrandom field γ(x , ω), e.g. φ(γ) = exp(γ).Expanding φ in a series in the Hermite polynomials:
φ(γ) =∞∑
i=0
φihi(γ), φi =
+∞∫−∞
φ(z)1i!
hi(z) exp(−z2/2)dz, (2)
where hi(z) is the i-th Hermite polynomial.[see PhD of E. Zander 2013, or PhD of A. Keese, 2005]
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
14 / 41
4*
Connection of cov. matrices for κ(x , ω) and γ(x , ω)
First, given the covariance matrix of κ(x , ω), we may relate itwith the covariance matrix of γ(x , ω) as follows,
covκ(x , y) =
∫(κ(x , ω)− κ(x)) (κ(y , ω)− κ(y)) dP(ω)
≈Q∑
i=0
i!φ2i covi
γ(x , y).
Solving this implicit Q-order equation [E. Zander, 13], we derivecovγ(x , y). Now, the KLE may be computed,
γ(x , ω) =∞∑
m=1
gm(x)θm(ω),
∫D
covγ(x , y)gm(y)dy = λmgm(x),
(3)
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
15 / 41
4*
Full JM,p and sparse J spM,p multi-index sets
DefinitionThe full multi-index is defined by restricting each componentindependently,
(αk )] is defined by rk−1nk rknumbers, where nk is number of grid points (e.g. nk = pk + 1)in the αk direction, and rk is the TT rank. The total number ofentries O(Mnr2), r = maxrk.
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
24 / 41
4*
Example: M-dimensional Laplacian
It has the Kronecker (canonical) rank-M representation:
with A = tridiag−1,2,−1 ∈ Rn×n, and I the n × n identity.In the TT format is explicitly representable with all TT ranksequal to 2:
A = (A I) 1(
I 0A I
)1 ... 1
(I 0A I
)1
(IA
), (16)
Or
A(i, j) =(A(i1, j1) I(i1, j1)
)( I(i2, j2) 0A(i2, j2) I(i2, j2)
)· · ·(
I(id , jd )A(id , jd )
).
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
25 / 41
4*
Low-rank response surface: PCE in the TT format
Calculation of
κα(`) =(α1 + · · ·+ αM)!
α1! · · ·αM !φα1+···+αM
∫D
M∏m=1
gαmm (x)v`(x)dx .
in tensor formats needs:I given a procedure to compute each element of a tensor,
e.g. κα1,...,αM by (26).I build a TT approximation κα ≈ κ(1)(α1) · · ·κ(M)(αM) using
a feasible amount of elements (i.e. much less than(p + 1)M ).
Such procedure exists, and relies on the cross interpolation ofmatrices, generalized to a higher-dimensional case [Oseledets,Tyrtyshnikov 2010; Savostyanov 13; Grasedyck; Bebendorf].
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
26 / 41
Skip 3 technical slides about Maximum volumeprinciple and its application
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
27 / 41
As soon as the reduced PCE coefficients κα(`) are computed,the initial expansion (8) comes easily. Indeed, stop the crossiteration at the first block, that is
κα(`) =∑
s1,...,sM−1
κ(1)`,s1
(α1) · · ·κ(M)sM−1
(αM). (17)
Now, collect the spatial components into the “zeroth” TT block,
κ(0)(x) =[κ(0)` (x)
]L
`=0=[κ(x) v1(x) · · · vL(x)
], (18)
then the PCE (4) writes as the following TT format,
κα(x) =∑
`,s1,...,sM−1
κ(0)` (x)κ
(1)`,s1
(α1) · · ·κ(M)sM−1
(αM). (19)
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
28 / 41
4*
Stochastic Galerkin matrix in TT format
Given (19), we split the whole sum over γ in (13):
∑γ∈JM,p
∆γ κγ(`) =∑
s1,...,sM−1
p∑γ1=0
∆γ1κ(1)`,s1
(γ1)
⊗· · ·⊗ p∑γM=0
∆γMκ(M)sM−1
(γM)
.
Introduce
K(0)(i , j) :=[K(0)` (i , j)
]L
`=0=[K0(i , j) K1(i , j) · · · KL(i , j)
], i , j = 1, . . . ,N,
K(m)sm−1,sm :=
∑pγm=0 ∆γmκ
(m)sm−1,sm (γm) for m = 1, . . . ,M,
then the TT representation for the operator writes
K =∑
`,s1,...,sM−1
K(0)` ⊗ K(1)
`,s1⊗ · · · ⊗ K(M)
sM−1∈ R(N·#JM,p)×(N·#JM,p), (20)
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
29 / 41
4*
Post-processing:
We computeCharacteristic, level sets, frequency in TT format
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
30 / 41
4*
Numerics: Main steps
1. Compute PCE of the coefficients κ(x , ω) in TT format2. Compute stochastic Galerkin matrix K in TT3. Compute solution of the linear system in TT4. Post-processing in TT format
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
31 / 41
4*
Numerics: Initial data and software
κ(x , ω) obeys the β5,2-distribution,covκ(x , y) = exp
(−(x − y)2/σ2) with σ = 0.3. D is L-shape
domain, 557 DOFs.Use sglib (E. Zander, TU BS) for discretization and solution withJ sp
M,p.Use TT-Toolbox for full JM,p.Use sglib for low-dimensional stages,and replace high-dimensional calculations by the TT.Use amen cross.m for TT approximation of κα (26),Use amen solve.m ( tAMEn, Dolgov) as linear system solverin TT format.
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
32 / 41
4*
Computation of the PCE for the permeability coefficient
Table : CPU times (sec.) of the permeability assembly
Center for Uncertainty Quantification Logo Lock-up
36 / 41
4*
Errors in the solution covariance matrices, | covu − cov?u |
The reference covariance matrix cov?u ∈ RN×N is computed inthe TT format with p = 5, and the discrepancies in the resultswith smaller p are calculated in average over all spatial points,
Center for Uncertainty Quantification Logo Lock-up
37 / 41
4*
Take to home
1. demonstrated RS in TT format for solving PDEs withuncertain coefficients.
2. Favor of the TT comparing to CP is a stable quasi-optimalrank reduction based on SVD.
3. Complexity O(Mnr3) with full accuracy control.4. TT methods become preferable for high p, but otherwise
the full computation in a small sparse set may be incrediblyfast. This reflects well the “curse of order”, taking place forthe sparse set instead of the “curse of dimensionality” inthe full set: the cardinality of the sparse set growsexponentially with p.
5. The TT approach scales linearly with p.
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
38 / 41
4*
Take to home
1. TT methods allow easy calculation of the stochasticGalerkin operator. With p below 10, the TT storage of theoperator allows us to forget about the sparsity issues,since the number of TT entries O(Mp2r2) is tractable.
2. Other polynomial families, such as the Chebyshev orLaguerre, may be incorporated into the scheme freely.
3. TT formalism may be recommended for stochastic PDEsas a general tool: one introduces the same discretizationlevels for all variables and let the algorithms determine aquasi-optimal representation adaptivity.
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
39 / 41
4*
Many questions are still open
1. Can we endow the solution scheme with more structureand obtain a more efficient algorithm?
2. Is there a better way to discretize stochastic fields than theKLE-PCE approach?
3. In the preliminary experiments, we have investigated onlythe simplest statistics, i.e. mean and variance. Whatquantities (level sets, frequency,...) are feasible in TTformat and how can they be effectively computed?
Center for UncertaintyQuantification
Center for UncertaintyQuantification
Center for Uncertainty Quantification Logo Lock-up
40 / 41
4*
Stochastic Galerkin library
1. Type in your terminalgit clone git://github.com/ezander/sglib.git